Revisit Lmser and its further development based on convolutional layers

  • 2019-04-12 16:26:04
  • Wenjing Huang, Shikui Tu, Lei Xu
  • 2


Proposed in 1991, Least Mean Square Error Reconstruction for self-organizingnetwork, shortly Lmser, was a further development of the traditionalauto-encoder (AE) by folding the architecture with respect to the centralcoding layer and thus leading to the features of symmetric weights and neurons,as well as jointly supervised and unsupervised learning. However, itsadvantages were only demonstrated in a one-hidden-layer implementation due tothe lack of computing resources and big data at that time. In this paper, werevisit Lmser from the perspective of deep learning, develop Lmser networkbased on multiple convolutional layers, which is more suitable forimage-related tasks, and confirm several Lmser functions with preliminarydemonstrations on image recognition, reconstruction, association recall, and soon. Experiments demonstrate that Lmser indeed works as indicated in theoriginal paper, and it has promising performance in various applications.


Quick Read (beta)

loading the full paper ...